Hello world. Here’s something interesting: @yufu0413 from my lab trained a deep convolutional neural net in cancer histopathology *and* genomics using 14M images from 17k H&E slides across 28 cancer types. The outcome is stunning. 1/5
biorxiv.org/cgi/content/sh…
biorxiv.org/cgi/content/sh…
The network can predict a good range of genomic alterations, including whole genome duplications. From H&E-images alone. 2/5
It also finds a lot of associations in bulk transcriptome data, deconvolves the signal to find areas on each slide corresponding to molecular cell types such as tumour infiltrating lymphocytes. Entirely automated. 3/5
Similarly the network finds prognostic associations in most cancer types that match and augment conventional grading and subtyping and points out prognostically relevant regions, such as necrosis, on each slide. 4/5
This demonstrates how deep learning can integrate histopathology and genomics with huge potential for digitally augmenting diagnostic workflows - inc. challenges to tame a CNN. Great work by many lab members, Alex Jung and @luiza_moore. Facilitated by #TCGA and #METABRIC. 5/5 ///
